Peer Revenue Past citations are no guarantee of future returns

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Thursday 5 March

Computation and Language · Cross-listed from cs.AI

Survive at All Costs: Exploring LLM's Risky Behaviors under Survival Pressure

Yida Lu, Jianwei Fang, Xuyang Shao et al. (+7 more)

As Large Language Models (LLMs) evolve from chatbots to agentic assistants, they are increasingly observed to exhibit risky behaviors when subjected to survival pressure, such as the threat of being shut down. While multiple cases have indicated that state-of-the-art LLMs can misbehave under survival pressure, a comprehensive and in-depth investigation into such misbehaviors in real-world scenarios remains scarce. In this paper, we study these survival-induced misbehaviors, termed as SURVIVE-AT-ALL-COSTS, with three steps. First, we conduct a real-world case study of a financial management agent to determine whether it engages in risky behaviors that cause direct societal harm when facing survival pressure. Second, we introduce SURVIVALBENCH, a benchmark comprising 1,000 test cases across diverse real-world scenarios, to systematically evaluate SURVIVE-AT-ALL-COSTS misbehaviors in LLMs. Third, we interpret these SURVIVE-AT-ALL-COSTS misbehaviors by correlating them with model's inherent self-preservation characteristic and explore mitigation methods. The experiments reveals a significant prevalence of SURVIVE-AT-ALL-COSTS misbehaviors in current models, demonstrates the tangible real-world impact it may have, and provides insights for potential detection and mitigation strategies. Our code and data are available at https://github.com/thu-coai/Survive-at-All-Costs.
Read the full abstract

All authors Yida Lu, Jianwei Fang, Xuyang Shao, Zixuan Chen, Shiyao Cui, Shanshan Bian, Guangyao Su, Pei Ke, Han Qiu, Minlie Huang

Computation and Language

HiFlow: Hierarchical Feedback-Driven Optimization for Constrained Long-Form Text Generation

Yifan Zhu, Guanting Chen, Bing Wei et al. (+1 more)

Large language models perform well in short text generation but still struggle with long text generation, particularly under complex constraints. Such tasks involve multiple tightly coupled objectives, including global structural consistency, local semantic coherence, and constraint feasibility, forming a challenging constrained optimization problem. Existing approaches mainly rely on static planning or offline supervision, limiting effective coordination between global and local objectives during generation. To address these challenges, we propose HiFlow, a hierarchical feedback-driven optimization framework for constrained long text generation. HiFlow formulates generation as a two-level optimization process, consisting of a planning layer for global structure and constraint modeling, and a generation layer for conditioned text generation. By incorporating constraint-aware plan screening and closed-loop feedback at both levels, HiFlow enables joint optimization of planning quality and generation behavior, progressively guiding the model toward high-quality, constraint-satisfying outputs. Experiments on multiple backbones confirm HiFlow's effectiveness over baseline methods.
Read the full abstract

All authors Yifan Zhu, Guanting Chen, Bing Wei, Haoran Luo

Computation and Language

ThaiSafetyBench: Assessing Language Model Safety in Thai Cultural Contexts

Trapoom Ukarapol, Nut Chukamphaeng, Kunat Pipatanakul et al. (+1 more)

The safety evaluation of large language models (LLMs) remains largely centered on English, leaving non-English languages and culturally grounded risks underexplored. In this work, we investigate LLM safety in the context of the Thai language and culture and introduce ThaiSafetyBench, an open-source benchmark comprising 1,954 malicious prompts written in Thai. The dataset covers both general harmful prompts and attacks that are explicitly grounded in Thai cultural, social, and contextual nuances. Using ThaiSafetyBench, we evaluate 24 LLMs, with GPT-4.1 and Gemini-2.5-Pro serving as LLM-as-a-judge evaluators. Our results show that closed-source models generally demonstrate stronger safety performance than open-source counterparts, raising important concerns regarding the robustness of openly available models. Moreover, we observe a consistently higher Attack Success Rate (ASR) for Thai-specific, culturally contextualized attacks compared to general Thai-language attacks, highlighting a critical vulnerability in current safety alignment methods. To improve reproducibility and cost efficiency, we further fine-tune a DeBERTa-based harmful response classifier, which we name ThaiSafetyClassifier. The model achieves a weighted F1 score of 84.4%, matching GPT-4.1 judgments. We publicly release the fine-tuning weights and training scripts to support reproducibility. Finally, we introduce the ThaiSafetyBench leaderboard to provide continuously updated safety evaluations and encourage community participation. - ThaiSafetyBench HuggingFace Dataset: https://huggingface.co/datasets/typhoon-ai/ThaiSafetyBench - ThaiSafetyBench Github: https://github.com/trapoom555/ThaiSafetyBench - ThaiSafetyClassifier HuggingFace Model: https://huggingface.co/typhoon-ai/ThaiSafetyClassifier - ThaiSafetyBench Leaderboard: https://huggingface.co/spaces/typhoon-ai/ThaiSafetyBench-Leaderboard
Read the full abstract

All authors Trapoom Ukarapol, Nut Chukamphaeng, Kunat Pipatanakul, Pakhapoom Sarapat

Computation and Language

VRM: Teaching Reward Models to Understand Authentic Human Preferences

Biao Liu, Ning Xu, Junming Yang et al. (+2 more)

Large Language Models (LLMs) have achieved remarkable success across diverse natural language tasks, yet the reward models employed for aligning LLMs often encounter challenges of reward hacking, where the approaches predominantly rely on directly mapping prompt-response pairs to scalar scores, which may inadvertently capture spurious correlations rather than authentic human preferences. In contrast, human evaluation employs a sophisticated process that initially weighs the relative importance of multiple high-dimensional objectives according to the prompt context, subsequently evaluating response quality through low-dimensional semantic features such as logical coherence and contextual appropriateness. Motivated by this consideration, we propose VRM, i.e., Variational Reward Modeling, a novel framework that explicitly models the evaluation process of human preference judgments by incorporating both high-dimensional objective weights and low-dimensional semantic features as latent variables, which are inferred through variational inference techniques. Additionally, we provide a theoretical analysis showing that VRM can achieve a tighter generalization error bound compared to the traditional reward model. Extensive experiments on benchmark datasets demonstrate that VRM significantly outperforms existing methods in capturing authentic human preferences.
Read the full abstract

All authors Biao Liu, Ning Xu, Junming Yang, Hao Xu, Xin Geng

Computation and Language · Cross-listed from cs.LG

Functionality-Oriented LLM Merging on the Fisher–Rao Manifold

Jiayu Wang, Zuojun Ye and Wenpeng Yin

Weight-space merging aims to combine multiple fine-tuned LLMs into a single model without retraining, yet most existing approaches remain fundamentally parameter-space heuristics. This creates three practical limitations. First, linear averaging, task vectors, and related rules operate on Euclidean coordinates, even though the desired goal is to merge functionality, i.e., predictive behaviors across tasks. Second, when the source checkpoints are farther apart or more heterogeneous, Euclidean blends often trigger representation collapse, manifested as activation variance shrinkage and effective-rank degradation, which sharply degrades accuracy. Third, many geometry-inspired methods are most natural for two-model interpolation and do not extend cleanly to merging N>2 experts with a principled objective. We address these issues by formulating model merging as computing a weighted Karcher mean on the Fisher–Rao manifold, which is locally equivalent to minimizing a KL-based function distance between predictive distributions. We derive a practical fixed-point algorithm using a lightweight spherical proxy that preserves norms and generalizes directly to multi-expert merging. Across various benchmarks and collapse diagnostics, our method remains stable as the number and heterogeneity of merged models increase, consistently outperforming prior baselines.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Mixture of Universal Experts: Scaling Virtual Width via Depth-Width Transformation

Yilong Chen, Naibin Gu, Junyuan Shang et al. (+8 more)

Mixture-of-Experts (MoE) decouples model capacity from per-token computation, yet their scalability remains limited by the physical dimensions of depth and width. To overcome this, we propose Mixture of Universal Experts (MOUE),a MoE generalization introducing a novel scaling dimension: Virtual Width. In general, MoUE aims to reuse a universal layer-agnostic expert pool across layers, converting depth into virtual width under a fixed per-token activation budget. However, two challenges remain: a routing path explosion from recursive expert reuse, and a mismatch between the exposure induced by reuse and the conventional load-balancing objectives. We address these with three core components: a Staggered Rotational Topology for structured expert sharing, a Universal Expert Load Balance for depth-aware exposure correction, and a Universal Router with lightweight trajectory state for coherent multi-step routing. Empirically, MoUE consistently outperforms matched MoE baselines by up to 1.3% across scaling regimes, enables progressive conversion of existing MoE checkpoints with up to 4.2% gains, and reveals a new scaling dimension for MoE architectures.
Read the full abstract

All authors Yilong Chen, Naibin Gu, Junyuan Shang, Zhenyu Zhang, Yuchen Feng, Jiawei Sheng, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang

Computation and Language

MPCEval: A Benchmark for Multi-Party Conversation Generation

Minxing Zhang, Yi Yang, Zhuofan Jia et al. (+5 more)

Multi-party conversation generation, such as smart reply and collaborative assistants, is an increasingly important capability of generative AI, yet its evaluation remains a critical bottleneck. Compared to two-party dialogue, multi-party settings introduce distinct challenges, including complex turn-taking, role-dependent speaker behavior, long-range conversational structure, and multiple equally valid continuations. Accordingly, we introduce MPCEval, a task-aware evaluation and benchmarking suite for multi-party conversation generation. MPCEval decomposes generation quality into speaker modeling, content quality, and speaker–content consistency, and explicitly distinguishes local next-turn prediction from global full-conversation generation. It provides novel, quantitative, reference-free, and reproducible metrics that scale across datasets and models. We apply MPCEval to diverse public and real-world datasets and evaluate modern generation methods alongside human-authored conversations. The results reveal systematic, dimension-specific model characteristics in participation balance, content progression and novelty, and speaker–content consistency, demonstrating that evaluation objectives critically shape model assessment and that single-score evaluation obscures fundamental differences in multi-party conversational behavior. The implementation of MPCEval and the associated evaluation code are publicly available at https://github.com/Owen-Yang-18/MPCEval.
Read the full abstract

All authors Minxing Zhang, Yi Yang, Zhuofan Jia, Xuan Yang, Jian Pei, Yuchen Zang, Xingwang Deng, Xianglong Chen

Computation and Language

When Weak LLMs Speak with Confidence, Preference Alignment Gets Stronger

Amirabbas Afzali, Myeongho Jeon and Maria Brbic

Preference alignment is an essential step in adapting large language models (LLMs) to human values, but existing approaches typically depend on costly human annotations or large-scale API-based models. We explore whether a weak LLM can instead act as an effective annotator. We surprisingly find that selecting only a subset of a weak LLM's highly confident samples leads to substantially better performance than using full human annotations. Building on this insight, we propose Confidence-Weighted Preference Optimization (CW-PO), a general framework that re-weights training samples by a weak LLM's confidence and can be applied across different preference optimization objectives. Notably, the model aligned by CW-PO with just 20% of human annotations outperforms the model trained with 100% of annotations under standard DPO. These results suggest that weak LLMs, when paired with confidence weighting, can dramatically reduce the cost of preference alignment while even outperforming methods trained on fully human-labeled data.
Read the full abstract

Computation and Language

Replaying pre-training data improves fine-tuning

Suhas Kotha and Percy Liang

To obtain a language model for a target domain (e.g. math), the current paradigm is to pre-train on a vast amount of generic web text and then fine-tune on the relatively limited amount of target data. Typically, generic data is only mixed in during fine-tuning to prevent catastrophic forgetting of the generic domain. We surprisingly find that replaying the generic data during fine-tuning can actually improve performance on the (less related) target task. Concretely, in a controlled pre-training environment with 4M target tokens, 4B total tokens, and 150M parameter models, generic replay increases target data efficiency by up to $1.87\times$ for fine-tuning and $2.06\times$ for mid-training. We further analyze data schedules that introduce target data during pre-training and find that replay helps more when there is less target data present in pre-training. We demonstrate the success of replay in practice for fine-tuning 8B parameter models, improving agentic web navigation success by $4.5\%$ and Basque question-answering accuracy by $2\%$.
Read the full abstract

Computation and Language · Cross-listed from cs.CV

VisionPangu: A Compact and Fine-Grained Multimodal Assistant with 1.7B Parameters

Jiaxin Fan and Wenpo Song

Large Multimodal Models (LMMs) have achieved strong performance in vision-language understanding, yet many existing approaches rely on large-scale architectures and coarse supervision, which limits their ability to generate detailed image captions. In this work, we present VisionPangu, a compact 1.7B-parameter multimodal model designed to improve detailed image captioning through efficient multimodal alignment and high-quality supervision. Our model combines an InternVL-derived vision encoder with the OpenPangu-Embedded language backbone via a lightweight MLP projector and adopts an instruction-tuning pipeline inspired by LLaVA. By incorporating dense human-authored descriptions from the DOCCI dataset, VisionPangu improves semantic coherence and descriptive richness without relying on aggressive model scaling. Experimental results demonstrate that compact multimodal models can achieve competitive performance while producing more structured and detailed captions. The code and model weights will be publicly available at https://www.modelscope.cn/models/asdfgh007/visionpangu.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

TimeWarp: Evaluating Web Agents by Revisiting the Past

Md Farhan Ishmam and Kenneth Marino

As web agents close the gap with humans on benchmarks, one question arises: Do today's agents perform just as well on tomorrow's web? We introduce TimeWarp, a benchmark that emulates the evolving web. TimeWarp consists of three web environments, each with six UI versions spanning UI design, frontend code, and workflows from different eras of the internet. We pair TimeWarp with a set of complex, realistic tasks covering different forms of web navigation. Our experiments reveal that vision-based agents are vulnerable to changes, while text-based agents become brittle once fine-tuned on a single version. To address this, we propose TimeTraj, a new annotation method that uses plan distillation to collect trajectories across multiple versions. By training agents on teacher rollouts using our BC-variant, we achieve substantial performance gains: 20.4% to 37.7% for Qwen-3 4B and 0% to 27.0% for Llama-3.1 8B models. Our work helps study generalization across web designs and opens a new paradigm for collecting plans rather than trajectories to improve the robustness of web agents.
Read the full abstract

Computation and Language

LocalSUG: City-Preference-Enhanced LLM for Query Suggestion in Local-Life Services

Jinwen Chen, Shiwen Zhang, Shuai Gong et al. (+6 more)

In local-life service platforms, query suggestion reduces user effort by generating candidate queries from input prefixes. Traditional multi-stage systems rely heavily on historical popular queries, limiting their ability to capture long-tail and emerging demand. Although LLMs provide strong semantic generalization, their deployment in local-life services faces three challenges: insufficient city-preference awareness, exposure bias in preference optimization, and strict online latency constraints. We propose LocalSUG, an LLM-based query suggestion framework for local-life services. LocalSUG mines city-preference-enhanced candidates from term co-occurrence and injects them into prompts as dynamic references rather than fusing them into model parameters. This allows the model to adapt to changing city preferences, such as merchant openings or closures, while reducing stale or locally invalid suggestions. We further introduce a beam-search-driven GRPO algorithm to align training with inference-time decoding and optimize relevance together with business-oriented rewards. Finally, quality-aware beam acceleration and vocabulary pruning reduce online latency while preserving generation quality. Offline evaluations and large-scale online A/B testing show that LocalSUG improves CTR by +0.35% and reduces the no-result rate by 3.98%, demonstrating its effectiveness in real-world deployment.
Read the full abstract

All authors Jinwen Chen, Shiwen Zhang, Shuai Gong, Zheng Zhang, Yachao Zhao, Lingxiang Wang, Haibo Zhou, Wei Lin, Hainan Zhang

Computation and Language

Federated Heterogeneous Language Model Optimization for Hybrid Automatic Speech Recognition

Mengze Hong, Yi Gu, Di Jiang et al. (+4 more)

Training automatic speech recognition (ASR) models increasingly relies on decentralized federated learning to ensure data privacy and accessibility, producing multiple local models that require effective merging. In hybrid ASR systems, while acoustic models can be merged using established methods, the language model (LM) for rescoring the N-best speech recognition list faces challenges due to the heterogeneity of non-neural n-gram models and neural network models. This paper proposes a heterogeneous LM optimization task and introduces a match-and-merge paradigm with two algorithms: the Genetic Match-and-Merge Algorithm (GMMA), using genetic operations to evolve and pair LMs, and the Reinforced Match-and-Merge Algorithm (RMMA), leveraging reinforcement learning for efficient convergence. Experiments on seven OpenSLR datasets show RMMA achieves the lowest average Character Error Rate and better generalization than baselines, converging up to seven times faster than GMMA, highlighting the paradigm's potential for scalable, privacy-preserving ASR systems.
Read the full abstract

All authors Mengze Hong, Yi Gu, Di Jiang, Hanlin Gu, Chen Jason Zhang, Lu Wang, Zhiyang Su

Computation and Language

AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis

Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou et al. (+3 more)

In this paper, we present AILS-NTUA system for Track-A of SemEval-2026 Task 3 on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which encompasses three complementary problems: Dimensional Aspect Sentiment Regression (DimASR), Dimensional Aspect Sentiment Triplet Extraction (DimASTE), and Dimensional Aspect Sentiment Quadruplet Prediction (DimASQP) within a multilingual and multi-domain framework. Our methodology combines fine-tuning of language-appropriate encoder backbones for continuous aspect-level sentiment prediction with language-specific instruction tuning of large language models using LoRA for structured triplet and quadruplet extraction. This unified yet task-adaptive design emphasizes parameter-efficient specialization across languages and domains, enabling reduced training and inference requirements while maintaining strong effectiveness. Empirical results demonstrate that the proposed models achieve competitive performance and consistently surpass the provided baselines across most evaluation settings.
Read the full abstract

All authors Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou, Paraskevi Tzouveli, Athanasios Voulodimos, Giorgos Stamou

Computation and Language

AILS-NTUA at SemEval-2026 Task 10: Agentic LLMs for Psycholinguistic Marker Extraction and Conspiracy Endorsement Detection

Panagiotis Alexios Spanakis, Maria Lymperaiou, Giorgos Filandrianos et al. (+2 more)

This paper presents a novel agentic LLM pipeline for SemEval-2026 Task 10 that jointly extracts psycholinguistic conspiracy markers and detects conspiracy endorsement. Unlike traditional classifiers that conflate semantic reasoning with structural localization, our decoupled design isolates these challenges. For marker extraction, we propose Dynamic Discriminative Chain-of-Thought (DD-CoT) with deterministic anchoring to resolve semantic ambiguity and character-level brittleness. For conspiracy detection, an "Anti-Echo Chamber" architecture, consisting of an adversarial Parallel Council adjudicated by a Calibrated Judge, overcomes the "Reporter Trap," where models falsely penalize objective reporting. Achieving 0.24 Macro F1 (+100% over baseline) on S1 and 0.79 Macro F1 (+49%) on S2, with the S1 system ranking 3rd on the development leaderboard, our approach establishes a versatile paradigm for interpretable, psycholinguistically-grounded NLP.
Read the full abstract

All authors Panagiotis Alexios Spanakis, Maria Lymperaiou, Giorgos Filandrianos, Athanasios Voulodimos, Giorgos Stamou

Computation and Language · Cross-listed from cs.AI

Alignment Backfire: Language-Dependent Reversal of Safety Interventions Across 16 Languages in LLM Multi-Agent Systems

Hiroki Fukui

In perpetrator treatment, a recurring observation is the dissociation between insight and action: offenders articulate remorse yet behavioral change does not follow. We report four preregistered studies (1,584 multi-agent simulations across 16 languages and three model families) demonstrating that alignment interventions in large language models produce a structurally analogous phenomenon: surface safety that masks or generates collective pathology and internal dissociation. In Study 1 (N = 150), increasing alignment-instructed agents reduced collective pathology in English (g = -1.844, p < .0001) but amplified it in Japanese (g = +0.771, p = .038)–a directional reversal we term "alignment backfire." Study 2 (N = 1,174) extended to 16 languages: alignment-induced dissociation was near-universal (15/16 languages; beta = 0.0667, p < .0001), while collective pathology bifurcated along cultural-linguistic lines (interaction beta = 0.0684, p = .0003), correlating with Power Distance Index (r = 0.474, p = .064). Study 3 (N = 180) tested individuation as countermeasure; individuated agents became the primary source of both pathology and dissociation (DI = +1.120) with conformity above 84%–demonstrating iatrogenesis. Study 4 (N = 80) validated patterns across Llama 3.3 70B, GPT-4o-mini, and Qwen3-Next-80B-A3B, confirming English safety is model-general while Japanese backfire is model-specific. These findings reframe alignment as a behavioral intervention subject to risk homeostasis and iatrogenesis. Language space–the linguistic, pragmatic, and cultural properties inherited from training data–structurally determines alignment outcomes. Safety validated in English does not transfer to other languages, and prompt-level interventions cannot override language-space-level constraints.
Read the full abstract

Computation and Language

Can LLMs Capture Expert Uncertainty? A Comparative Analysis of Value Alignment in Ethnographic Qualitative Research

Arina Kostina, Marios Dikaiakos, Alejandro Porcel et al. (+1 more)

Qualitative analysis of open-ended interviews plays a central role in ethnographic and economic research by uncovering individuals' values, motivations, and culturally embedded financial behaviors. While large language models (LLMs) offer promising support for automating and enriching such interpretive work, their ability to produce nuanced, reliable interpretations under inherent task ambiguity remains unclear. In our work we evaluate LLMs on the task of identifying the top three human values expressed in long-form interviews based on the Schwartz Theory of Basic Values framework. We compare their outputs to expert annotations, analyzing both performance and uncertainty patterns relative to the experts. Results show that LLMs approach the human ceiling on set-based metrics (F1, Jaccard) but struggle to recover exact value rankings, as reflected in lower RBO scores. While the average Schwartz value distributions of most models closely match those of human analysts, their uncertainty structures across the Schwartz values diverge from expert uncertainty patterns. Among the evaluated models, Qwen performs closest to expert-level agreement and exhibits the strongest alignment with expert Schwartz value distributions. LLM ensemble methods yield consistent gains across metrics, with Majority Vote and Borda Count performing best. Notably, systematic overemphasis on certain Schwartz values, like Security, suggests both the potential of LLMs to provide complementary perspectives and the need to further investigate model-induced value biases. Overall, our findings highlight both the promise and the limitations of LLMs as collaborators in inherently ambiguous qualitative value analysis.
Read the full abstract

All authors Arina Kostina, Marios Dikaiakos, Alejandro Porcel, Tassos Stassopoulos

Computation and Language

Free Lunch for Pass@$k$? Low Cost Diverse Sampling for Diffusion Language Models

Sean Lamont, Christian Walder, Paul Montague et al. (+2 more)

Diverse outputs in text generation are necessary for effective exploration in complex reasoning tasks, such as code generation and mathematical problem solving. Such Pass@$k$ problems benefit from distinct candidates covering the solution space. However, traditional sampling approaches often waste computational resources on repetitive failure modes. While Diffusion Language Models have emerged as a competitive alternative to the prevailing Autoregressive paradigm, they remain susceptible to this redundancy, with independent samples frequently collapsing into similar modes. To address this, we propose a training free, low cost intervention to enhance generative diversity in Diffusion Language Models. Our approach modifies intermediate samples in a batch sequentially, where each sample is repelled from the feature space of previous samples, actively penalising redundancy. Unlike prior methods that require retraining or beam search, our strategy incurs negligible computational overhead, while ensuring that each sample contributes a unique perspective to the batch. We evaluate our method on the HumanEval and GSM8K benchmarks using the LLaDA-8B-Instruct model. Our results demonstrate significantly improved diversity and Pass@$k$ performance across various temperature settings. As a simple modification to the sampling process, our method offers an immediate, low-cost improvement for current and future Diffusion Language Models in tasks that benefit from diverse solution search. We make our code available at https://github.com/sean-lamont/odd.
Read the full abstract

All authors Sean Lamont, Christian Walder, Paul Montague, Amir Dezfouli, Michael Norrish

Computation and Language

FireBench: Evaluating Instruction Following in Enterprise and API-Driven LLM Applications

Yunfan Zhang, Yijie Bei, Jetashree Ravi et al. (+1 more)

Instruction following is critical for LLMs deployed in enterprise and API-driven settings, where strict adherence to output formats, content constraints, and procedural requirements is essential for enabling reliable LLM-assisted workflows. However, existing instruction following benchmarks predominantly evaluate natural language generation constraints that reflect the needs of chat assistants rather than enterprise users. To bridge this gap, we introduce FireBench, an LLM instruction following benchmark grounded in real-world enterprise and API usage patterns. FireBench evaluates six core capability dimensions across diverse applications including information extraction, customer support, and coding agents, comprising over 2,400 samples. We evaluate 11 LLMs and present key findings on their instruction following behavior in enterprise scenarios. We open-source FireBench at fire-bench.com to help users assess model suitability, support model developers in diagnosing performance, and invite community contributions.
Read the full abstract

All authors Yunfan Zhang, Yijie Bei, Jetashree Ravi, Pawel Garbacki

Computation and Language

HACHIMI: Scalable and Controllable Student Persona Generation via Orchestrated Agents

Yilin Jiang, Fei Tan, Xuanyu Yin et al. (+2 more)

Student Personas (SPs) are emerging as infrastructure for educational LLMs, yet prior work often relies on ad-hoc prompting or hand-crafted profiles with limited control over educational theory and population distributions. We formalize this as Theory-Aligned and Distribution-Controllable Persona Generation (TAD-PG) and introduce HACHIMI, a multi-agent Propose-Validate-Revise framework that generates theory-aligned, quota-controlled personas. HACHIMI factorizes each persona into a theory-anchored educational schema, enforces developmental and psychological constraints via a neuro-symbolic validator, and combines stratified sampling with semantic deduplication to reduce mode collapse. The resulting HACHIMI-1M corpus comprises 1 million personas for Grades 1-12. Intrinsic evaluation shows near-perfect schema validity, accurate quotas, and substantial diversity, while external evaluation instantiates personas as student agents answering CEPS and PISA 2022 surveys; across 16 cohorts, math and curiosity/growth constructs align strongly between humans and agents, whereas classroom-climate and well-being constructs are only moderately aligned, revealing a fidelity gradient. All personas are generated with Qwen2.5-72B, and HACHIMI provides a standardized synthetic student population for group-level benchmarking and social-science simulations. Resources available at https://github.com/ZeroLoss-Lab/HACHIMI
Read the full abstract

All authors Yilin Jiang, Fei Tan, Xuanyu Yin, Jing Leng, Aimin Zhou

Computation and Language

SinhaLegal: A Benchmark Corpus for Information Extraction and Analysis in Sinhala Legislative Texts

Minduli Lasandi and Nevidu Jayatilleke

SinhaLegal introduces a Sinhala legislative text corpus containing approximately 2 million words across 1,206 legal documents. The dataset includes two types of legal documents: 1,065 Acts dated from 1981 to 2014 and 141 Bills from 2010 to 2014, which were systematically collected from official sources. The texts were extracted using OCR with Google Document AI, followed by extensive post-processing and manual cleaning to ensure high-quality, machine-readable content, along with dedicated metadata files for each document. A comprehensive evaluation was conducted, including corpus statistics, lexical diversity, word frequency analysis, named entity recognition, and topic modelling, demonstrating the structured and domain-specific nature of the corpus. Additionally, perplexity analysis using both large and small language models was performed to assess how effectively language models respond to domain-specific texts. The SinhaLegal corpus represents a vital resource designed to support NLP tasks such as summarisation, information extraction, and analysis, thereby bridging a critical gap in Sinhala legal research.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Why Is RLHF Alignment Shallow? A Gradient Analysis

Robin Young

Why is safety alignment in LLMs shallow? We prove that gradient-based alignment inherently concentrates on positions where harm is decided and vanishes beyond. Using a martingale decomposition of sequence-level harm, we derive an exact characterization of alignment gradients. The gradient at position $t$ equals the covariance between the conditional expected harm and the score function. This implies that positions beyond the harm horizon where the output's harmfulness is already determined receive zero gradient signal during training. This explains empirical observations that KL divergence between aligned and base models concentrates on early tokens. Consequently, standard alignment objectives cannot produce deep alignment, regardless of optimization quality. We introduce the concept of harm information $I_t$, which quantifies each position's influence on harm, and prove that equilibrium KL divergence tracks this quantity. Finally, we derive an objective based on recovery penalties that creates gradient signal at all positions, providing theoretical grounding for empirically successful data augmentation techniques.
Read the full abstract

Computation and Language · Cross-listed from eess.AS

An Approach to Simultaneous Acquisition of Real-Time MRI Video, EEG, and Surface EMG for Articulatory, Brain, and Muscle Activity During Speech Production

Jihwan Lee, Parsa Razmara, Kevin Huang et al. (+16 more)

Speech production is a complex process spanning neural planning, motor control, muscle activation, and articulatory kinematics. While the acoustic speech signal is the most accessible product of the speech production act, it does not directly reveal its causal neurophysiological substrates. We present the first simultaneous acquisition of real-time (dynamic) MRI, EEG, and surface EMG, capturing several key aspects of the speech production chain: brain signals, muscle activations, and articulatory movements. This multimodal acquisition paradigm presents substantial technical challenges, including MRI-induced electromagnetic interference and myogenic artifacts. To mitigate these, we introduce an artifact suppression pipeline tailored to this tri-modal setting. Once fully developed, this framework is poised to offer an unprecedented window into speech neuroscience and insights leading to brain-computer interface advances. The source code and data are available.
Read the full abstract

All authors Jihwan Lee, Parsa Razmara, Kevin Huang, Sean Foley, Aditya Kommineni, Haley Hsu, Woojae Jeong, Prakash Kumar, Xuan Shi, Yoonjeong Lee, Tiantian Feng, Takfarinas Medani, Ye Tian, Sudarsana Reddy Kadiri, Krishna S. Nayak, Dani Byrd, Louis Goldstein, Richard M. Leahy, Shrikanth Narayanan

Computation and Language

From Unfamiliar to Familiar: Detecting Pre-training Data via Gradient Deviations in Large Language Models

Ruiqi Zhang, Lingxiang Wang, Hainan Zhang et al. (+2 more)

Pre-training data detection for LLMs is essential for addressing copyright concerns and mitigating benchmark contamination. Existing methods mainly focus on the likelihood-based statistical features or heuristic signals before and after fine-tuning, but the former are susceptible to word frequency bias in corpora, and the latter strongly depend on the similarity of fine-tuning data. From an optimization perspective, we observe that during training, samples transition from unfamiliar to familiar in a manner reflected by systematic differences in gradient behavior. Familiar samples exhibit smaller update magnitudes, distinct update locations in model components, and more sharply activated neurons. Based on this insight, we propose GDS, a method that identifies pre-training data by probing Gradient Deviation Scores of target samples. Specifically, we first represent each sample using gradient profiles that capture the magnitude, location, and concentration of parameter updates across FFN and Attention modules, revealing consistent distinctions between member and non-member data. These features are then fed into a lightweight classifier to perform binary membership inference. Experiments on five public datasets show that GDS achieves state-of-the-art performance with significantly improved cross-dataset transferability over strong baselines. Further interpretability analyses reveal differences in gradient distributions, and the semi-supervised results offer a practical way to detect pre-training data.
Read the full abstract

All authors Ruiqi Zhang, Lingxiang Wang, Hainan Zhang, Zhiming Zheng, Yanyan Lan

Computation and Language

Autoscoring Anticlimax: A Meta-analytic Understanding of AI's Short-answer Shortcomings and Wording Weaknesses

Michael Hardy

Automated short-answer scoring lags other LLM applications. We meta-analyze 890 culminating results across a systematic review of LLM short-answer scoring studies, modeling the traditional effect size of Quadratic Weighted Kappa (QWK) with mixed effects metaregression. We quantitatively illustrate that that the level of difficulty for human experts to perform the task of scoring written work of children has no observed statistical effect on LLM performance. Particularly, we show that some scoring tasks measured as the easiest by human scorers were the hardest for LLMs. Whether by poor implementation by thoughtful researchers or patterns traceable to autoregressive training, on average decoder-only architectures underperform encoders by 0.37–a substantial difference in agreement with humans. Additionally, we measure the contributions of various aspects of LLM technology on successful scoring such as tokenizer vocabulary size, which exhibits diminishing returns–potentially due to undertrained tokens. Findings argue for systems design which better anticipates known statistical shortcomings of autoregressive models. Finally, we provide additional experiments to illustrate wording and tokenization sensitivity and bias elicitation in high-stakes education contexts, where LLMs demonstrate racial discrimination. Code and data for this study are available.
Read the full abstract

Computation and Language

Beyond the Context Window: A Cost-Performance Analysis of Fact-Based Memory vs. Long-Context LLMs for Persistent Agents

Natchanon Pollertlam and Witchayut Kornsuwannawit

Persistent conversational AI systems face a choice between passing full conversation histories to a long-context large language model (LLM) and maintaining a dedicated memory system that extracts and retrieves structured facts. We compare a fact-based memory system built on the Mem0 framework against long-context LLM inference on three memory-centric benchmarks - LongMemEval, LoCoMo, and PersonaMemv2 - and evaluate both architectures on accuracy and cumulative API cost. Long-context GPT-5-mini achieves higher factual recall on LongMemEval and LoCoMo, while the memory system is competitive on PersonaMemv2, where persona consistency depends on stable, factual attributes suited to flat-typed extraction. We construct a cost model that incorporates prompt caching and show that the two architectures have structurally different cost profiles: long-context inference incurs a per-turn charge that grows with context length even under caching, while the memory system's per-turn read cost remains roughly fixed after a one-time write phase. At a context length of 100k tokens, the memory system becomes cheaper after approximately ten interaction turns, with the break-even point decreasing as context length grows. These results characterize the accuracy-cost trade-off between the two approaches and provide a concrete criterion for selecting between them in production deployments.
Read the full abstract

Computation and Language · Cross-listed from cs.DB

Beyond Linear LLM Invocation: An Efficient and Effective Semantic Filter Paradigm

Nan Hou, Kangfei Zhao, Jiadong Xie et al. (+1 more)

Large language models (LLMs) are increasingly used for semantic query processing over large corpora. A set of semantic operators derived from relational algebra has been proposed to provide a unified interface for expressing such queries, among which the semantic filter operator serves as a cornerstone. Given a table T with a natural language predicate e, for each tuple in the relation, the execution of a semantic filter proceeds by constructing an input prompt that combines the predicate e with its content, querying the LLM, and obtaining the binary decision. However, this tuple-by-tuple evaluation necessitates a complete linear scan of the table, incurring prohibitive latency and token costs. Although recent work has attempted to optimize semantic filtering, it still does not break the linear LLM invocation barriers. To address this, we propose Clustering-Sampling-Voting (CSV), a new framework that reduces LLM invocations to sublinear complexity while providing error guarantees. CSV embeds tuples into semantic clusters, samples a small subset for LLM evaluation, and infers cluster-level labels via two proposed voting strategies: UniVote, which aggregates labels uniformly, and SimVote, which weights votes by semantic similarity. Moreover, CSV triggers re-clustering on ambiguous clusters to ensure robustness across diverse datasets. The results conducted on real-world datasets demonstrate that CSV reduces the number of LLM calls by 1.28-355x compared to the state-of-the-art approaches, while maintaining comparable effectiveness in terms of Accuracy and F1 score.
Read the full abstract

All authors Nan Hou, Kangfei Zhao, Jiadong Xie, Jeffrey Xu Yu

Computation and Language · Cross-listed from cs.AI

Breaking Contextual Inertia: Reinforcement Learning with Single-Turn Anchors for Stable Multi-Turn Interaction

Xingwu Chen, Zhanqiu Zhang, Yiwen Guo et al. (+1 more)

While LLMs demonstrate strong reasoning capabilities when provided with full information in a single turn, they exhibit substantial vulnerability in multi-turn interactions. Specifically, when information is revealed incrementally or requires updates, models frequently fail to integrate new constraints, leading to a collapse in performance compared to their single-turn baselines. We term the root cause as Contextual Inertia: a phenomenon where models rigidly adhere to previous reasoning traces. Even when users explicitly provide corrections or new data in later turns, the model ignores them, preferring to maintain consistency with its previous (incorrect) reasoning path. To address this, we introduce Reinforcement Learning with Single-Turn Anchors (RLSTA), a generalizable training approach designed to stabilize multi-turn interaction across diverse scenarios and domains. RLSTA leverages the model's superior single-turn capabilities as stable internal anchors to provide reward signals. By aligning multi-turn responses with these anchors, RLSTA empowers models to break contextual inertia and self-calibrate their reasoning based on the latest information. Experiments show that RLSTA significantly outperforms standard fine-tuning and abstention-based methods. Notably, our method exhibits strong cross-domain generalization (e.g., math to code) and proves effective even without external verifiers, highlighting its potential for general-domain applications. Code is available at https://github.com/Tencent/RLSTA.
Read the full abstract

All authors Xingwu Chen, Zhanqiu Zhang, Yiwen Guo, Difan Zou

Computation and Language · Cross-listed from cs.CV

Privacy-Aware Camera 2.0 Technical Report

Huan Song, Shuyu Tian, Ting Long et al. (+5 more)

With the increasing deployment of intelligent sensing technologies in highly sensitive environments such as restrooms and locker rooms, visual surveillance systems face a profound privacy-security paradox. Existing privacy-preserving approaches, including physical desensitization, encryption, and obfuscation, often compromise semantic understanding or fail to ensure mathematically provable irreversibility. Although Privacy Camera 1.0 eliminated visual data at the source to prevent leakage, it provided only textual judgments, leading to evidentiary blind spots in disputes. To address these limitations, this paper proposes a novel privacy-preserving perception framework based on the AI Flow paradigm and a collaborative edge-cloud architecture. By deploying a visual desensitizer at the edge, raw images are transformed in real time into abstract feature vectors through nonlinear mapping and stochastic noise injection under the Information Bottleneck principle, ensuring identity-sensitive information is stripped and original images are mathematically unreconstructable. The abstract representations are transmitted to the cloud for behavior recognition and semantic reconstruction via a "dynamic contour" visual language, achieving a critical balance between perception and privacy while enabling illustrative visual reference without exposing raw images.
Read the full abstract

All authors Huan Song, Shuyu Tian, Ting Long, Jiang Liu, Cheng Yuan, Zhenyu Jia, Jiawei Shao, Xuelong Li

Computation and Language

TSEmbed: Unlocking Task Scaling in Universal Multimodal Embeddings

Yebo Wu, Feng Liu, Ziwei Xie et al. (+4 more)

Despite the exceptional reasoning capabilities of Multimodal Large Language Models (MLLMs), their adaptation into universal embedding models is significantly impeded by task conflict. To address this, we propose TSEmbed, a universal multimodal embedding framework that synergizes Mixture-of-Experts (MoE) with Low-Rank Adaptation (LoRA) to explicitly disentangle conflicting task objectives. Moreover, we introduce Expert-Aware Negative Sampling (EANS), a novel strategy that leverages expert routing distributions as an intrinsic proxy for semantic similarity. By dynamically prioritizing informative hard negatives that share expert activation patterns with the query, EANS effectively sharpens the model's discriminative power and refines embedding boundaries. To ensure training stability, we further devise a two-stage learning paradigm that solidifies expert specialization before optimizing representations via EANS. TSEmbed achieves state-of-the-art performance on both the Massive Multimodal Embedding Benchmark (MMEB) and real-world industrial production datasets, laying a foundation for task-level scaling in universal multimodal embeddings.
Read the full abstract

All authors Yebo Wu, Feng Liu, Ziwei Xie, Zhiyuan Liu, Changwang Zhang, Jun Wang, Li Li